基于图神经网络的水富—宜宾航道多站点水位预报模型

Multi station water level prediction model for Shuifu to Yibin waterway based on graph neural network

受到岷江、横江影响,向家坝下游水富—宜宾段水位变化特性复杂,干支流间水位、流量数据属于多维时空数据。研究选取spectral temporal graph neural network(StemGNN)时空图神经网络用于向家坝下游多站点水位预报,结果表明:该方法适用于研究区域的多站点水位预报,未来1、8 h模型预报性能较优,在向家坝站、宜宾站、李庄站3处的最大预报误差约为0.5 m。StemGNN特点是能够从输入数据中自动提取河网结构信息,体现研究区域的汇流特性。横江流量对于研究区域水位流量影响较小;向家坝水库水位、横江水位、高场水位代表研究区域前期的水位情况,高场流量作为较大的流量输入,对于研究区域水位流量影响较大。研究成果可为近坝段、支流入汇等水位变化特性复杂河段的多站点水位预报提供新思路。

Affected by Minjiang River and Hengjiang River,the water level variation characteristics of Shuifu-Yibin section in the lower reaches of Xiangjiaba are complex.The water level and flow data between the main and branch streams belong to multi-dimensional spatio-temporal data.The study selects spectral temporal graph neural network(StemGNN)for multi station water level prediction in the downstream of Xiangjiaba.The results show that the method is suitable for multi station water level prediction in the study area,and the prediction performance of the model in the next 1 hour and next 8 hours is better.The maximum prediction error at Xiangjiaba station,Yibin station and Lizhuang station is about 0.5 m.The characteristics of StemGNN are that it can automatically extract the river network structure information from the input data,reflecting the confluence characteristics of the study area.gao_yulei@ctg.com.cn Hengjiang River flow has little impact on the water level and flow in the study area;Xiangjiaba reservoir water level,Hengjiang water level and Gaochang water level represent the early stage water level of the study area.Gaochang flow,as a large flow input,has a great impact on the water level and flow of the study area.The research results can provide new ideas for multi station water level prediction in the river reach with complex water level variation characteristics,such as near dam section and tributary inflow.